Ensuring secure communication in wireless networks remains a significant challenge, especially in the presence of untrusted relays. This study addresses this challenge by employing an intelligent Transmit Antenna Selection (TAS) technique, integrated with machine learning (ML) and deep learning (DL) models, including the Transformer architecture, to optimize secure transmission over Rician fading channels with hardware impairments (HWIs). Simulation results demonstrate that the Transformer-based model achieves state-of-the-art performance, consistently surpassing traditional ML models in classification accuracy and security metrics. Specifically, the Transformer achieves an average Area Under the Curve (AUC) of 0.90 and classification accuracy of 83.5%, significantly outperforming the Support Vector Machine (SVM) (AUC: 0.87, accuracy: 41.5%), k-nearest Neighbors (KNN) (AUC: 0.69, accuracy: 46%), and Naive Bayes (NB) (AUC: 0.53, accuracy: 18%). Furthermore, the Transformer model approximates the performance of exhaustive search methods in terms of average secrecy rate and secrecy outage probability, highlighting its ability to capture complex dependencies and adapt to diverse scenarios. These findings establish the Transformer as a robust and scalable solution for secure communication in challenging wireless environments, paving the way for future advancements in physical layer security (PLS) and intelligent TAS strategies.